Figure 6 - uploaded by Jonathan How
Content may be subject to copyright.
Download
View publication
Increase of conservatism with . An agent (orange) following the CARRL policy avoids a dynamic, non-cooperative obstacle (blue) that is observed without noise. An increasing robustness parameter (left to right) increases the agent's conservatism, i.e., the agent avoids the obstacle with a greater safety distance.

Increase of conservatism with . An agent (orange) following the CARRL policy avoids a dynamic, non-cooperative obstacle (blue) that is observed without noise. An increasing robustness parameter (left to right) increases the agent's conservatism, i.e., the agent avoids the obstacle with a greater safety distance.

Source publication
Figure 1: Intuition. An adversary distorts the true position, s0, of a...
Figure 6: Increase of conservatism with . An agent (orange) following...
Certified Adversarial Robustness for Deep Reinforcement Learning
Preprint
Full-text available
  • Oct 2019
Deep Neural Network-based systems are now the state-of-the-art in many robotics tasks, but their application in safety-critical domains remains dangerous without formal guarantees on network robustness. Small perturbations to sensor inputs (from noise or adversarial examples) are often enough to change network-based decisions, which was already sho...
Cite
Download full-text

Similar publications

Fig. 1. MAPE-K loop.
Fig. 2. Template to instantiate security assurance case for NIST...
Fig. 3. Instance of security control network for access control (partial).
Fig. 4. AC-3 instantiated within the SAC GSN template.
+6 Fig. 5. Expanded AC-3 module goals to operational goals needed to...
MAPE-K/MAPE-SAC: An interaction framework for adaptive systems with security assurance cases
Article
Full-text available
  • Mar 2020
Security certification establishes that a given system satisfies properties and constraints as specified in the system security profile. Mechanisms and techniques have been developed to assess if and how well the system complies with the properties, thereby providing a degree of confidence in the security certification. Generally, certification of...
View
Fig. 1: Intuition. An adversary distorts the true position, s0, of a...
Fig. 3: Illustration of -Ball, Bp(s, ), for n = 2. Let = [1, 2]....
Certified Adversarial Robustness for Deep Reinforcement Learning
Preprint
Full-text available
  • Apr 2020
Deep Neural Network-based systems are now the state-of-the-art in many robotics tasks, but their application in safety-critical domains remains dangerous without formal guarantees on network robustness. Small perturbations to sensor inputs (from noise or adversarial examples) are often enough to change network-based decisions, which was recently sh...
View